CEM_SAC

Source code for paper "Combining Soft-Actor Critic with Cross-Entropy Method for Policy Search in Continuous Control". IEEE CEC 2022

Hieu Trung Nguyen, Khang Tran and Ngoc Hoang Luong

Setup

• Clone this repo:

$ git clone https://github.com/junhill-2000/eTD3.git
$ cd CEC_SAC

• The following packages are needed:

tianshou==0.4.2
gym==0.19.0
mujoco_py==2.0.2.13 (need to install mujoco200 from http://www.mujoco.org/ or try my sh script)

Install Old Version of Mujoco

gdown --id 1H6PkDQGlp5bUfvanbYPv-Of9ZZ8OSDHJ
chmod +x mujoco_colab.sh 
./mujoco_colab.sh

import os
os.environ['LD_LIBRARY_PATH'] += ':'
os.environ['LD_LIBRARY_PATH'] += '/root/.mujoco/mujoco200/bin'
os.environ.get("LD_LIBRARY_PATH", "")

Usage

train agent with CEM_SAC

python3 cem_sac.py --task [environment_name] --seed [seed] --logdir [log_directory]

Google Colab Notebook

Try this notebook to run our code without local installation

Visualize (comming soon)

Acknowledgement

Our source code is inspired by:

• CEM-RL by Pourchot et al.
• Tianshou library